# LLM Generation Efficiency on Qwen 2.5 14B (2048 input + 256 generation tokens)

Source: [LLM Generation Efficiency on Qwen 2.5 14B (2048 input + 256 generation tokens)](<https://www.sota2.com/research/sota/llm-generation-efficiency-on-qwen-2-5-14b-2048-input-plus-256-generation-tokens>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T05:42:00.784Z

Evaluation of the generation efficiency metrics (End-to-end Latency and Time to Next Token) for the Qwen 2.5 14B model using a context window of 2048 input tokens and generating 256 output tokens.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | LLM Generation Efficiency |
| Dataset (including any stated split) | Qwen 2.5 14B (2048 input + 256 generation tokens) |
| Metric | End-to-end Latency (s) |
| Unit | TIME |
| Direction | Lower is better |
| Results loaded | 4 of 4 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [End-to-end Latency (s)](<https://www.sota2.com/research/sota/llm-generation-efficiency-on-qwen-2-5-14b-2048-input-plus-256-generation-tokens.md?metric=End-to-end%20Latency%20(s)>): Lower is better; unit: TIME.

- [Avg. Time to Next Token (ms)](<https://www.sota2.com/research/sota/llm-generation-efficiency-on-qwen-2-5-14b-2048-input-plus-256-generation-tokens.md?metric=Avg.%20Time%20to%20Next%20Token%20(ms)>): Lower is better; unit: TIME.

## Recorded results

| Method | End-to-end Latency (s) | Result date | Report paper |
| --- | --- | --- | --- |
| GRIFFIN | 8.7 | Not recorded | 68fa7bc51e209131cea8a58e |
| Caprese | 8.7 | Not recorded | 68fa7bc51e209131cea8a58e |
| LoRA | 9.5 | Not recorded | 68fa7bc51e209131cea8a58e |
| Full | 10.5 | Not recorded | 68fa7bc51e209131cea8a58e |

### GRIFFIN

End-to-end Latency (s): 8.7. Result ID: 699a867a5698665ec41077d9.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc51e209131cea8a58e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74701e209131cea61cb9.md>)

Recorded evaluation configuration: {"Sparsity":"50%","r":"256","Hardware":"NVIDIA L40 GPU","Precision":"BF16","Model Portion":"First half of the model"}. This record does not assert independent reproduction.

### Caprese

End-to-end Latency (s): 8.7. Result ID: 699a867a5698665ec41077db.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc51e209131cea8a58e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bc51e209131cea8a58e.md>)

Recorded evaluation configuration: {"Sparsity":"50%","r":"256","Hardware":"NVIDIA L40 GPU","Precision":"BF16"}. This record does not assert independent reproduction.

### LoRA

End-to-end Latency (s): 9.5. Result ID: 699a867a5698665ec41077da.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc51e209131cea8a58e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b121e209131cea2bf59.md>)

[Implementation](<https://github.com/microsoft/LoRA>)

Recorded evaluation configuration: {"Sparsity":"50%","r":"256","Hardware":"NVIDIA L40 GPU","Precision":"BF16","Configuration":"GRIFFIN paired with LoRA weights"}. This record does not assert independent reproduction.

### Full

End-to-end Latency (s): 10.5. Result ID: 699a867a5698665ec41077d8.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bc51e209131cea8a58e.md>)

Recorded evaluation configuration: {"Hardware":"NVIDIA L40 GPU","Precision":"BF16"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.